[Platform] Replace torch.cuda.Event with torch.Event (#47140)

Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Kunshang Ji
2026-06-30 08:39:59 -07:00
committed by GitHub
co-authored by Harry Mellor
parent fcaa84efa7
commit e840f0d3f5
31 changed files with 60 additions and 70 deletions
+2 -5
View File
@@ -272,7 +272,7 @@ class DeepseekV4Attention(nn.Module, AttentionLayerBase, ABC):
# [0]: GEMM start / post-GEMM event0. [1..3]: GEMM done events;
# [1] doubles as post-GEMM event1. Reuse is safe: GEMM fully joins
# before post-GEMM starts.
self.ln_events = [torch.cuda.Event() for _ in range(4)]
self.ln_events = [torch.Event() for _ in range(4)]
assert cache_config is not None, "DeepseekV4 attention requires cache_config"
# ---- Attention / KV-cache setup ----
@@ -760,10 +760,7 @@ class DeepseekV4Indexer(nn.Module):
# None on ROCm — maybe_execute_in_parallel falls back to sequential.
self.aux_stream = aux_stream
self.ln_events: list[torch.cuda.Event] = [
torch.cuda.Event(),
torch.cuda.Event(),
]
self.ln_events: list[torch.Event] = [torch.Event(), torch.Event()]
def forward(
self,
-11
View File
@@ -44,17 +44,6 @@ class DeepseekV4XPUAttention(DeepseekV4Attention):
backend_cls = DeepseekV4XPUSparseBackend
use_flashmla_fp8_layout = True
def __init__(self, *args, **kwargs) -> None:
# torch.cuda.Event() raises RuntimeError on XPU ("dummy base class").
# The Base and DeepseekV4Indexer both create cuda Events in __init__, so
# we temporarily redirect torch.cuda.Event → torch.xpu.Event.
_orig_event = torch.cuda.Event
torch.cuda.Event = torch.xpu.Event # type: ignore[misc]
try:
super().__init__(*args, **kwargs)
finally:
torch.cuda.Event = _orig_event # type: ignore[misc]
def _fused_qnorm_rope_kv_insert(self, q, kv, positions, attn_metadata):
from typing import cast